An AI feedback loop connects generated material, ratings, clicks, retrieved documents, or real-world outcomes back into future generation or selection. The loop can improve adaptation when feedback reflects genuine quality, but it can also amplify errors and weak proxies.
Research citation loops are dangerous because several model outputs may appear to be independent confirmations while ultimately repeating one unsupported origin. Provenance, deduplication, primary-source checks, and independent evaluation are needed to break the cycle.
ELI5
A feedback loop happens when the result of an AI process influences what the system does later. Helpful feedback can improve future choices, while misleading feedback can cause the same error to grow stronger over time.
For example, a recommendation system may show more of the videos people click. If clicks reflect genuine interest, the suggestions can improve, but sensational recommendations can also generate clicks and then become overrepresented.
